Datasets:
File size: 1,520 Bytes
6c1c833 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 | {
"field_id": "F047",
"field": "content moderation",
"directions": [
{
"direction_id": "G0",
"name": "Recognition of harmful content across evidence and settings",
"scope": "Develop harmful-content recognition capabilities and their annotation foundations, encompassing text detection, annotated corpora and schemes, adaptation across languages, cultures, and data distributions, visual or cross-modal judgments, and the use of conversational, community, or targeted-group context to identify abuse and hate."
},
{
"direction_id": "G1",
"name": "Evaluation of moderation-detector failures",
"scope": "Determine when, how, or for whom existing moderation detectors fail through targeted tests, adversarial inputs, distribution shifts, error analysis, or performance disparities."
},
{
"direction_id": "G2",
"name": "Situated and reason-grounded moderation judgments",
"scope": "Study how moderation judgments vary by judge, culture, or definition of harm, and how to explain the reasons for judgments using textual evidence, implicit biases, social effects, or moderation policies. These branches do not require both modeling disagreement and generating explanations."
},
{
"direction_id": "G3",
"name": "Generative interventions against harmful speech",
"scope": "Generate counterspeech, counter-narratives, or less harmful rewrites, including resources established primarily to support such outputs."
}
]
}
|